Researchers use computer simulations, AI to speed up drug development

Published July 24, 2024, by Judith Van Dongen

Every year, the FDA approves a few dozen new drugs. The approval marks the conclusion of a drug development process that can take up to 15 years and costs an estimated $2.6 billion for a single drug. The hefty price tag of bringing new drugs to market includes the cost of failed drugs; for every drug that makes it, thousands of others get shelved.

Reducing the time and cost to develop new drugs has been front of mind for Senthil Natesan, an associate professor in the College of Pharmacy and Pharmaceutical Sciences. Natesan’s expertise is in computer-aided drug design, which uses high-performance computing resources and techniques to study proteins involved in various diseases. It also identifies potential drug molecules to target those proteins. The approach allows scientists to use computer simulations to predict how a drug will interact with different proteins and how safe, effective, and selective that drug will be.

“These in silico, or computer-based, predictions significantly decrease the cost and time involved in assessing new drug compounds,” Natesan said. “For example, we can use these methods to predict the degree to which a drug molecule can pass through to the brain or reach its intended target when taken by mouth. Predicting such properties for millions of compounds using experimental methods would take several years, whereas we can predict them in silico in hours.”

As a result, computer-aided drug design can help expedite the introduction of new drugs. First conceived in the 1980s, the field played a significant role in the development of drugs to treat HIV and AIDS, Natesan said. Since then, it has accelerated based on knowledge gained from the Human Genome Project combined with advances in structural biology, bioinformatics, and computer technology. The most recent advances in the field come from generative artificial intelligence (AI), which can generate data by analyzing patterns in existing data.

In a recently published study in the Journal of Chemical Theory and Computation, Natesan and his team used a generative AI model to make predictions related to the interaction between drug molecules and proteins in the cell membrane. They found that their novel method cut down significantly on the time it takes to make these predictions.

“Our paper demonstrates the power of AI in obtaining the membrane partitioning profile of compounds in only a third of the typical computational time,” he said. “Using a high-performance workstation, it usually takes about a month to obtain that information for a single drug. Now we can get this data in just 10 days.”

The membrane separates a cell’s interior from its outside environment and regulates the transport of materials—such as drug molecules—that enter and exit the cell. Close to a third of all proteins in the human body reside in and around the cell membrane, and more than half of all FDA-approved drugs target these membrane proteins. Data on the interactions between drug molecules and membrane proteins helps scientists evaluate how drugs are transported across organs, tissues, and cells, which gives them valuable information on a drug’s potency, onset and duration of action, and undesirable side effects.

Obtaining this information through experimental methods is expensive and time-consuming, which is why scientists have been using computational methods to predict these characteristics. In their study, Natesan and his research team were the first to combine existing computational methods with an AI model to speed up the process even more.

Rather than compute this information across the full length of the molecule layers in which these membrane proteins are embedded—which would require 60 calculations—Natesan and his team used their novel method to compute the information for 10 selected regions, fed those data to the AI model, and had it generate the missing data for the remaining 50 regions. Testing their model on existing data for three commonly used anti-asthma drugs and 20 other compounds, they showed that the model’s predictions were accurate.

Natesan said the potential impact of their work is immediate. He and his coauthors—including former graduate student Peter Obi, former postdoc Jeevan GC, current postdoc Charles Mariasoosai, and current graduate student Ayobami Diyaolu—have made the code for their model available on GitHub, where it can be accessed by pharmaceutical companies interested in implementing their method.

“Time is a crucial bottleneck in understanding the interactions of drugs with membranes, which is a critical step in designing drugs that target membrane proteins,” Natesan said. “With this new methodology, we have successfully addressed this problem.”